Google unveiled "Gemini 1.5 Pro," an updated version of its multimodal artificial intelligence (AI) model "Gemini 1.0 Pro," to the world on the 15th (local time). Jeminai 1.5 Pro is a medium-sized model that generates and understands various types of information, including text, images, audio and video, and is said to have similar performance to Google's latest model, Gemini1.0 Ultra.
The main feature of Jeminai 1.5 Pro is that the amount of information that can be processed simultaneously is much higher than that of existing models. This is called the context window, and Jeminai 1.5 Pro has a huge context window of up to 1 million tokens. A token refers to the work of cutting a sentence into detailed units for AI to process, which is a unit with meanings such as words, images, videos, audio, codes, etc.
This means that Gemini 1.5 Pro can simultaneously process information equivalent to more than 700,000 words, 30,000 lines of code, 1 hour of video, and 11 hours of voice. Gemini 1.5 Pro's ability to understand and analyze such long contexts increases scalability and efficiency for a wide range of tasks.
"Gemini 1.5 Pro delivers dramatically improved performance," said Demis Hassabis, CEO of Google DeepMind. He explained that the development of the model represents a step change in the company's approach, which is based on the evolution of its underlying models and research and engineering innovations in almost every part of its infrastructure. One of these is the new Mixture-of-Experts (MoE) architecture, which is said to have increased the learning and service efficiency of Gemini 1.5.
Gemini 1.5 Pro will be available as a private preview to a limited group of developers and enterprise customers starting May 15, with up to 1 million token context windows available through AI Studio and Vertex AI. AI Studio is the fastest way to build a Gemini model and allows developers to easily integrate the Gemini API into their applications. Vertex AI is a platform where enterprises can deploy AI models. Gemini 1.5 Pro is available in 38 languages, including Korean, in more than 180 countries and regions, including Korea.
Gemini 1.5 Pro also has the ability to infer multiple amounts of information. This means that a large amount of content within a given query can be seamlessly analyzed, categorized, and summarized. For example, given a 402-page transcript of the Apollo 11 lunar mission, Gemini 1.5 Pro can infer each of the conversations, events, and details found throughout the document.
In addition, with improved comprehension and multiple inference capabilities, Gemini 1.5 Pro can perform highly sophisticated comprehension and inference tasks for multiple modalities, including video. For example, given a Buster Keaton movie, which is a 44-minute silent film, Gemini 1.5 Pro is able to accurately analyze different plots and events, and infer even small details that are easily missed in the movie.
Gemini 1.5 Pro can also perform more relevant debugging tasks on longer blocks of code. When prompted with more than 100,000 lines of code, Gemini 1.5 Pro can make better inferences through examples, suggest useful fixes, and provide explanations of how different parts of the code work.
When tested on a comprehensive panel of text, code, image, audio, and video evaluations, Google reported that Gemini 1.5 Pro outperformed Jeminai 1.0 Pro on 87 percent of the benchmarks used to develop Large Language Models (LLMs). It also significantly outperforms Jeminai 1.0 Ultra on the same benchmark.
Gemini 1.5 Pro maintains a high level of performance despite the large context window. In the Needle In A Haystack (NIAH) evaluation, which deliberately places small pieces of text containing specific facts or statements within long blocks of text, we found the contained text with 99% probability in data blocks of up to 1 million tokens.
Gemini 1.5 Pro also demonstrated impressive in-context learning capabilities, enabling it to perform well on tasks such as translating information it had never seen before.
All in all, Gemini 1.5 Pro combines the latest technology with high performance and is expected to make human ability and knowledge equal in a variety of fields.
